AN EXPLAINABLE DEEP LEARNING FRAMEWORK FOR BREAST CANCER CLASSIFICATION USING EFFICIENTNETV2B0 AND GRAD-CAM

Abstract

Breast cancer remains one of the most serious health challenges worldwide, withearly and accurate diagnosis can significantly improve patient outcomes. Traditional diagnostic methods often rely heavily on expert interpretation, which may lead to inconsistencies or delays in decision-making. To address this issue, this research provides a deep-learning framework that uses the EfficientNetV2B0 model in combination with Grad-CAM (Gradient-weighted Class Activation Mapping) to provide illustrated explanations to detect breast cancer using ultrasound and MRI datasets. Our method addresses serious challenges such as class imbalance and irrelevant image characteristics by employing SMOTE (Synthetic Minority Over-sampling Technique) oversampling and Region of Interest (ROI) extraction for BUSI (Breast Ultrasound Images) datasets. The Grad-CAM approach improves reliability and transparency by providing visual proof that support’s each decision, allowing healthcare professionals to better understand the AI's prediction. Trained and assessed on two different medical imaging datasets, the framework obtained extraordinarily high accuracy (98.97% on BUSI and 99.55% on MRI), along with low prediction error and high reliability. The model is both accurate and understandable, making it ideal for clinical usage. It is also faster and more dependable than current approaches, making it highly beneficial.